KR20210085278A - Apparatus and method for learning imbalanced data - Google Patents

Abstract

Disclosed are a device and a method for learning imbalanced data. According to an embodiment of the present invention, an unbalanced data learning device includes: one or more processors; and an executable memory for storing at least one or more programs to be executed by the one or more processors, wherein the at least one program receives imbalanced data, determines a weight based on the imbalanced data, and learns the imbalanced data using a predefined loss function in which a loss value changes based on the weight. Therefore, by learning the imbalanced data, a recognition rate degradation problem which occurs in a process of performing multi-class classification is solved.

Description

Unbalanced data learning apparatus and method {APPARATUS AND METHOD FOR LEARNING IMBALANCED DATA}

The present invention relates to machine learning technology, and more particularly to unbalanced data learning technology.

Supervised learning-based deep neural networks have recently shown strong performance based on their high representation power. This is because generalization is possible through network parameters by learning a lot of data. Therefore, a sufficient amount of data becomes the basis for learning, and correct label information is important for accurate network performance.

However, it is not easy to secure a sufficient amount of data for every class depending on the characteristics of the data domain. For example, in the case of an expression recognition database for six facial expression classifications collected from the web, you can find a lot of data of the Happiness and Sadness classes, but the Fear, Disgust, Surprise, and Anger emotion class facial expression images are the essence of the class. Due to the characteristics, the number of samples is small and it is relatively difficult to collect. Although it is possible to collect data in a balanced way by artificially making facial expressions in a laboratory environment, it is known that the generalization ability is poor, and performance is generally poor in real-world scenarios. As a simple way to solve the problem, you can balance it by adjusting the amount of data in other classes based on a small amount of data in one class. However, this is not a simple problem because it causes a problem in which the overall amount of data decreases and the classification performance of the deep neural network decreases.

Meanwhile, U.S. Patent No. 9,224,104 “Generating data from imbalanced training data sets” discloses a method of generating unbalanced training data with unbalanced label distribution as balanced data.

An object of the present invention is to solve the problem of lowering the recognition rate occurring in the process of performing multi-class classification by learning imbalanced data.

An apparatus for learning imbalanced data according to an embodiment of the present invention for achieving the above object includes one or more processors and an execution memory for storing at least one or more programs executed by the one or more processors, the at least one program may receive imbalance data, determine a weight based on the imbalance data, and learn the imbalance data using a predefined loss function in which a loss value is changed based on the weight.

In this case, the loss function may be used as an objective function for a BACK-PROPAGATION algorithm.

In this case, the at least one program may determine a weight based on at least one of a data distribution for each class of the imbalanced data and a recognition accuracy for each class for each batch of the imbalanced data (BATCH).

At this time, the at least one or more programs

In order to determine the weight based on the data distribution for each class, the weight may be determined so that the loss value is changed based on the amount occupied by the data of the class for each class.

In this case, the at least one program determines the weight in a direction in which the loss value increases when the amount of data of the class for each class is less than a preset value, and the amount of data of the class for each class is If it is greater than the set value, the weight may be determined in a direction in which the loss value is decreased.

In this case, the at least one program may calculate a hit rate for each class in order to determine a weight based on the recognition accuracy for each class.

In this case, the at least one program may increase the weight more than the preset weight if the hit rate is less than a preset value, and decrease the weight than the preset weight if the hit rate is higher than the preset value. .

At this time, the at least one program recrystallizes the weight for each batch based on the changed loss value and the weight, and calculates the loss function for each batch using the recrystallized weight to learn the imbalance data. can

In addition, the unbalanced data learning method according to an embodiment of the present invention for achieving the above object is the unbalanced data learning method of the unbalanced data learning apparatus, the step of receiving unbalanced data and determining a weight based on the unbalanced data and learning the imbalanced data using a predefined loss function in which a loss value changes based on the weight.

In this case, the loss function may be used as an objective function for a BACK-PROPAGATION algorithm.

In this case, in the learning of the imbalance data, the weight may be determined based on at least one of a data distribution for each class of the imbalance data and a recognition accuracy for each class for each batch of the imbalance data.

In this case, the learning of the imbalance data may determine the weight so that the loss value is changed based on the amount occupied by the data of the class for each class in order to determine the weight based on the data distribution for each class.

At this time, the step of learning the imbalance data determines the weight in a direction in which the loss value increases when the amount of data of the class for each class is less than a preset value, and the amount of data of the class for each class If it is greater than this preset value, the weight may be determined in a direction in which the loss value is decreased.

In this case, the learning of the imbalance data may calculate a hit rate for each class in order to determine a weight based on the recognition accuracy for each class.

In this case, the learning of the imbalance data may include increasing the weight than the predetermined weight if the hit rate is less than a predetermined value, and decreasing the weight than the predetermined weight if the hit rate is higher than the predetermined value. can

At this time, the step of learning the imbalance data comprises recrystallizing the weight for each batch based on the changed loss value and the weight, and calculating the loss function for each batch using the recrystallized weight to obtain the imbalance data. can learn

The present invention can solve the problem of lowering the recognition rate that occurs in the process of performing multi-class classification by learning imbalanced data.

1 is a block diagram illustrating an apparatus for learning imbalanced data according to an embodiment of the present invention.
2 is an operation flowchart illustrating a method for learning imbalanced data according to an embodiment of the present invention.
3 is a diagram illustrating a computer system according to an embodiment of the present invention.

The present invention will be described in detail with reference to the accompanying drawings as follows. Here, repeated descriptions, well-known functions that may unnecessarily obscure the gist of the present invention, and detailed descriptions of configurations will be omitted. The embodiments of the present invention are provided in order to more completely explain the present invention to those of ordinary skill in the art. Accordingly, the shapes and sizes of elements in the drawings may be exaggerated for clearer description.

Throughout the specification, when a part "includes" a certain element, it means that other elements may be further included, rather than excluding other elements, unless otherwise stated.

Hereinafter, preferred embodiments according to the present invention will be described in detail with reference to the accompanying drawings.

1 is a block diagram illustrating an apparatus for learning imbalanced data according to an embodiment of the present invention.

Referring to FIG. 1 , an apparatus for learning imbalanced data according to an embodiment of the present invention may include a data input unit 110 , a data learning unit 120 , a data output unit 130 , and a database unit 140 .

The data input unit 110 may receive IMBALANCED DATA for machine learning.

The data learning unit 120 may receive unbalanced data, determine a weight based on the unbalanced data, and learn the unbalanced data using a predefined loss function in which a loss value is changed based on the weight. .

In this case, the loss function may be used as an objective function for a BACK-PROPAGATION algorithm.

The data learning unit 120 may learn data using a loss function used in a deep neural network (DNN) structure.

The loss function can be used as an object function for the back-propagation algorithm when the DNN learns data.

For example, the loss function can be used in general DNN structures such as ConvolutionalNN and ResNet.

A cross entropy loss according to an embodiment of the present invention may be defined as in Equation (1).

는 학습 샘플,

는 그에 따른 레이블,

는 딥 뉴럴 네트워크의 파라메터를 나타낸다.

는 DNN을 나타내며 마지막 레이어는 소프트맥스(softmax) 함수이다. 여기서

의 정의는 수학식 2와 같이 나타낼 수 있다.When learning a total of n samples having c classes in Equation 1,

is the learning sample,

is the corresponding label,

denotes a parameter of a deep neural network.

represents the DNN and the last layer is a softmax function. here

The definition of can be expressed as Equation (2).

The data learner 120 may use a cross entropy function as a loss value by adding all values obtained by performing a -log operation to each class prediction probability.

함수에 의해 그 활성화 여부가 결정될 수 있다.At this time, the predicted probability for each class is

Whether or not it is activated can be determined by a function.

함수가 학습 데이터의 레이블 빈도 수에 비례하여 활성화될 수 있다.At this time, the data learning unit 120 performs learning while

A function may be activated in proportion to the number of label frequencies in the training data.

At this time, the data learning unit 120 uses an adaptive cross entropy loss function, which is a loss function according to an embodiment of the present invention, to overcome the disadvantage of cross entropy imbalance data can learn

The loss function according to an embodiment of the present invention can be expressed as Equation (3).

는 j번째 클래스의 샘플 개수를 나타내며,

는 해당 j번째 클래스의 hit 개수를 나타낸다.In Equation 3,

represents the number of samples in the jth class,

represents the number of hits of the j-th class.

In this case, the data learning unit 120 may determine a weight based on at least one of a data distribution for each class of the imbalanced data and a recognition accuracy for each class for each batch of the imbalanced data BATCH.

In this case, in order to determine the weight based on the data distribution for each class, the data learning unit 120 may determine the weight so that the loss value is changed based on the amount occupied by the data of the class for each class.

At this time, the data learning unit 120 determines the weight in a direction in which the loss value increases when the amount of data of the class for each class is less than a preset value, and the amount of data of the class for each class is If it is greater than a preset value, the weight may be determined in a direction in which the loss value is decreased.

항의 j번째 클래스의 데이터가 차지하는 양이 적으면 손실 값을 높여주는 방향으로 작동할 수 있다.At this time, the data learning unit 120

If the amount of data in the j-th class of the term is small, it can operate in the direction of increasing the loss value.

Conversely, when the amount of data occupied by the data learning unit 120 is large, the data learning unit 120 may operate in the direction of lowering the loss value, thereby mitigating the class-specific asymmetry of the loss value due to data imbalance.

Here, the c/(c-1) term may correspond to a constant term for normalizing the weight to a probability value.

In addition, in order to improve the accuracy for each class according to the data imbalance, the data learner 120 may calculate a hit rate for each class (HIT RATE) to determine a weight based on the recognition accuracy for each class.

In this case, the data learning unit 120 may increase the weight more than the preset weight if the hit rate is less than the preset value, and decrease the weight than the preset weight if the hit rate is higher than the preset value. have.

를 계산하며 클래스 별 히트 레이트(hit rate)가 낮으면 가중치를 높여주고, 반대로 높으면 가중치를 낮게 해주어, 예측 에러를 최소화함과 동시에 데이터 분포에 따른 정확도의 균형을 맞춰 줌으로써 전체적 인식률을 향상시킬 수 있다.At this time, the data learning unit 120 is a hit rate for each class during learning,

If the hit rate for each class is low, the weight is raised, and if the hit rate is high, the weight is lowered, thereby minimizing prediction error and balancing the accuracy according to the data distribution, thereby improving the overall recognition rate. .

When learning the DNN, the data learning unit 120 may calculate the loss function for each batch.

In Equation 3, n may be defined as the size of a batch.

When the data learning unit 120 defines a loss function for each batch, statistical information (accuracy for each class and data distribution for each class) may be defined in the loss function for each batch.

In this case, the data learning unit 120 may operate the backpropagation algorithm in a direction to alleviate the imbalanced data using the loss function defined for each batch.

As a result, the data learning unit 120 determines the weight of the cross entropy loss function for each batch based on the label statistical information for each batch (accuracy for each class and data distribution for each class), and as a result, the weight for each batch The loss function can be calculated by changing it.

At this time, the data learning unit 120 recrystallizes the weight for each batch based on the changed loss value and the weight, and calculates the loss function for each batch using the recrystallized weight to learn the imbalance data. can do.

The data output unit 130 may output a result of learning the imbalance data.

The database unit 140 may store a result of learning the imbalanced data.

2 is an operation flowchart illustrating a method for learning imbalanced data according to an embodiment of the present invention.

Referring to FIG. 2 , the method for learning imbalanced data according to an embodiment of the present invention may first receive data (S210).

That is, step S210 may receive IMBALANCED DATA for machine learning.

In addition, the method for learning imbalanced data according to an embodiment of the present invention can learn data (S220).

That is, in step S220, the imbalance data is received, a weight is determined based on the imbalance data, and the imbalance data can be learned using a predefined loss function in which a loss value is changed based on the weight. .

In step S220, data may be learned using a loss function used in a deep neural network (DNN) structure.

The loss function can be used as an object function for the back-propagation algorithm when the DNN learns data.

For example, the loss function can be used in general DNN structures such as ConvolutionalNN and ResNet.

A cross entropy loss according to an embodiment of the present invention may be defined as in Equation (1).

는 학습 샘플,

는 그에 따른 레이블,

는 딥 뉴럴 네트워크의 파라메터를 나타낸다.

는 DNN을 나타내며 마지막 레이어는 소프트맥스(softmax) 함수이다. 여기서

의 정의는 수학식 2와 같이 나타낼 수 있다.When learning a total of n samples having c classes in Equation 1,

is the learning sample,

is the corresponding label,

denotes a parameter of a deep neural network.

represents the DNN and the last layer is a softmax function. here

The definition of can be expressed as Equation (2).

Step S220 can be used as a loss value by adding all the values obtained by -log operation to each class prediction probability using a cross entropy function.

함수에 의해 그 활성화 여부가 결정될 수 있다.At this time, the predicted probability for each class is

Whether or not it is activated can be determined by a function.

함수가 학습 데이터의 레이블 빈도 수에 비례하여 활성화될 수 있다.At this time, step S220 is performed while learning

A function may be activated in proportion to the number of label frequencies in the training data.

At this time, step S220 is to overcome the disadvantage of cross entropy (Cross Entropy), learning imbalance data using the adaptive cross entropy loss function (Adaptive Cross Entropy Loss Function), which is a loss function according to an embodiment of the present invention can do.

The loss function according to an embodiment of the present invention can be expressed as Equation (3).

는 j번째 클래스의 샘플 개수를 나타내며,

는 해당 j번째 클래스의 hit 개수를 나타낸다.In Equation 3,

represents the number of samples in the jth class,

represents the number of hits of the j-th class.

In this case, in step S220 , the weight may be determined based on at least one of a data distribution for each class of the imbalanced data and a recognition accuracy for each class for each batch of the imbalanced data BATCH.

In this case, in step S220 , in order to determine the weight based on the data distribution for each class, the weight may be determined so that the loss value is changed based on the amount occupied by the data of the class for each class.

In this case, in step S220, if the amount of data of the class for each class is less than a preset value, the weight is determined in a direction in which the loss value increases, and the amount of data of the class for each class is a preset value. If greater, the weight may be determined in a direction in which the loss value is decreased.

항의 j번째 클래스의 데이터가 차지하는 양이 적으면 손실 값을 높여주는 방향으로 작동할 수 있다.At this time, step S220 is

If the amount of data in the j-th class of the term is small, it can operate in the direction of increasing the loss value.

Conversely, step S220 may operate in the direction of lowering the loss value when the amount of data occupied is large, and may operate in the direction of alleviating the class-specific asymmetry of the loss value due to data imbalance.

Here, the c/(c-1) term may correspond to a constant term for normalizing the weight to a probability value.

Also, in step S220, a hit rate for each class may be calculated to determine a weight based on the recognition accuracy for each class.

In this case, in step S220, if the hit rate is less than the preset value, the weight may be increased than the preset weight, and if the hit rate is higher than the preset value, the weight may be decreased than the preset weight.

를 계산하며 클래스별 히트 레이트(hit rate)가 낮으면 가중치를 높여주고, 반대로 높으면 가중치를 낮게 해주어, 예측 에러를 최소화함과 동시에 데이터 분포에 따른 정확도의 균형을 맞춰 줌으로써 전체적 인식률을 향상시킬 수 있다.Step S220 is a hit rate for each class during learning,

If the hit rate for each class is low, the weight is increased, and if the hit rate is high, the weight is decreased. .

At this time, in step S220, when training the DNN, the loss function may be calculated for each batch.

In Equation 3, n may be defined as the size of a batch.

In this case, in step S220, the backpropagation algorithm may be operated in a direction to alleviate the imbalanced data using the loss function defined for each batch.

As a result, in step S220, the weight of the cross entropy loss function for each batch is determined based on the label statistical information for each batch (accuracy for each class and data distribution for each class). As a result, by changing the weight for each batch The loss function can be calculated.

In this case, in step S220, the weight is reset for each batch based on the statistical information, and the loss function is calculated for each batch using the reset weight to learn the imbalance data.

In addition, the method for learning imbalanced data according to an embodiment of the present invention may output and store data (S230).

That is, step S230 may output and store the result of learning the imbalance data.

3 is a diagram illustrating a computer system according to an embodiment of the present invention.

Referring to FIG. 3 , the apparatus for learning imbalanced data according to an embodiment of the present invention may be implemented in a computer system 1100 such as a computer-readable recording medium. As shown in FIG. 3 , the computer system 1100 includes one or more processors 1110 , a memory 1130 , a user interface input device 1140 , and a user interface output device 1150 that communicate with each other via a bus 1120 . and storage 1160 . In addition, the computer system 1100 may further include a network interface 1170 coupled to the network 1180 . The processor 1110 may be a central processing unit or a semiconductor device that executes processing instructions stored in the memory 1130 or the storage 1160 . The memory 1130 and the storage 1160 may be various types of volatile or non-volatile storage media. For example, the memory may include a ROM 1131 or a RAM 1132 .

An apparatus for learning imbalanced data according to an embodiment of the present invention includes one or more processors 1110; and an execution memory 1130 for storing at least one or more programs executed by the one or more processors 1110, wherein the at least one program receives imbalance data, and determines a weight based on the imbalance data, , it is possible to learn the imbalance data using a predefined loss function in which the loss value changes based on the weight.

In this case, the loss function may be used as an objective function for a BACK-PROPAGATION algorithm.

In this case, the weight of the loss function may be set based on the hit rate (HIT RATE) of the class.

In this case, the at least one program may determine a weight based on at least one of a data distribution for each class of the imbalanced data and a recognition accuracy for each class for each batch of the imbalanced data (BATCH).

In this case, in order to determine the weight based on the data distribution for each class, the at least one program may determine the weight so that the loss value is changed based on the amount occupied by the data of the class for each class.

In this case, the at least one program determines the weight in a direction in which the loss value increases when the amount of data of the class for each class is less than a preset value, and the amount of data of the class for each class is If it is greater than the set value, the weight may be determined in a direction in which the loss value is decreased.

In this case, the at least one program may calculate a hit rate for each class in order to determine a weight based on the recognition accuracy for each class.

In this case, the at least one program may increase the weight more than the preset weight if the hit rate is less than a preset value, and decrease the weight than the preset weight if the hit rate is higher than the preset value. .

At this time, the at least one program recrystallizes the weight for each batch based on the changed loss value and the weight, and calculates the loss function for each batch using the recrystallized weight to learn the imbalance data. can

In the present invention, an adaptive loss function of a deep neural network is proposed to improve classification performance of data having data imbalance characteristics. The proposed adaptive loss function according to an embodiment of the present invention is configured in such a way that a high weight is given to a small amount of class data by reflecting the imbalance of data in the weight of the loss function, and the unbalanced class data through the proposed loss function Accurate classification results can be expected even with data with

As described above, in the apparatus and method for learning imbalanced data according to an embodiment of the present invention, the configuration and method of the embodiments described above are not limitedly applicable, but the embodiments are so that various modifications can be made. All or part of each embodiment may be selectively combined and configured.

110: data input unit 120: data learning unit
130: data output unit 140: database unit
1100: computer system 1110: processor
1120: bus 1130: memory
1131: rom 1132: ram
1140: user interface input device
1150: user interface output device
1160: storage 1170: network interface
1180: network

Claims (16)

one or more processors; and
an execution memory for storing at least one or more programs executed by the one or more processors;
including,
the at least one program
receiving unbalanced data,
Determining a weight based on the imbalance data, and learning the imbalance data by using a predefined loss function in which a loss value changes based on the weight. The method according to claim 1,
The loss function is
Unbalanced data learning apparatus, characterized in that it is used as an objective function for the BACK-PROPAGATION algorithm. 3. The method according to claim 2,
the at least one program
For each batch of the imbalanced data (BATCH), the imbalanced data learning apparatus, characterized in that the weight is determined based on at least one of a data distribution for each class of the imbalanced data and a recognition accuracy for each class. 4. The method according to claim 3,
the at least one program
In order to determine the weight based on the data distribution for each class, the weight is determined so that the loss value is changed based on the amount occupied by the data of the class for each class. 5. The method according to claim 4,
the at least one program
If the amount of data of the class for each class is smaller than a preset value, the weight is determined in a direction in which the loss value increases,
The apparatus for learning imbalanced data, characterized in that the weight is determined in a direction in which the loss value is decreased when the amount of data of each class is greater than a preset value. 4. The method according to claim 3,
the at least one program
In order to determine the weight based on the recognition accuracy for each class, the apparatus for learning unbalanced data, characterized in that calculating a hit rate (HIT RATE) for each class. 7. The method of claim 6,
the at least one program
When the hit rate is less than a preset value, the weight is increased than the preset weight, and when the hit rate is higher than the preset value, the weight is decreased than the preset weight. 4. The method according to claim 3,
the at least one program
Unbalanced data learning apparatus, characterized in that the weight is recrystallized for each batch based on the changed loss value and the weight, and the loss function is calculated for each batch using the recrystallized weight to learn the imbalance data. In the unbalanced data learning method of the unbalanced data learning apparatus,
receiving unbalanced data as input; and
determining a weight based on the imbalance data, and learning the imbalance data using a predefined loss function in which a loss value is changed based on the weight;
Unbalanced data learning method comprising a. 10. The method of claim 9,
The loss function is
Unbalanced data learning method, characterized in that it is used as an objective function for the BACK-PROPAGATION algorithm. 11. The method of claim 10,
The step of learning the imbalanced data is
For each batch of the imbalanced data (BATCH), the imbalanced data learning method, characterized in that the weight is determined based on at least one of a data distribution for each class of the imbalanced data and a recognition accuracy for each class. 12. The method of claim 11,
The step of learning the imbalanced data is
In order to determine the weight based on the data distribution for each class, the weight is determined so that the loss value is changed based on the amount occupied by the data of the class for each class. 13. The method of claim 12,
The step of learning the imbalanced data is
If the amount of data of the class for each class is smaller than a preset value, the weight is determined in a direction in which the loss value increases,
Unbalanced data learning method, characterized in that the weight is determined in a direction in which the loss value is reduced when the amount of data of the class for each class is greater than a preset value. 12. The method of claim 11,
The step of learning the imbalanced data is
In order to determine the weight based on the recognition accuracy for each class, the method for learning unbalanced data, characterized in that calculating a hit rate (HIT RATE) for each class. 15. The method of claim 14,
The step of learning the imbalanced data is
When the hit rate is less than a preset value, the weight is increased than the preset weight, and when the hit rate is higher than the preset value, the weight is decreased than the preset weight. 12. The method of claim 11,
The step of learning the imbalanced data is
Unbalanced data learning method, characterized in that the weight is re-determined for each batch based on the changed loss value and the weight, and the loss function is calculated for each batch using the re-determined weight to learn the imbalance data.

Images